Attachments

Attachments are structural, not text: what you pick is a chip on the composer’s pending strip, and it rides the next prompt without a mention ever appearing in the buffer. The finder is /attach [dir] — a fuzzy popup over the workspace (or over the directory you named) where typing narrows the list,

Enter/Tab attaches the highlighted file, and Ctrl+D changes the directory being searched without closing the dialog. [`/manage-attachments`](/docs/usage/commands) then edits the strip as one list: add another file through the same finder, reorder what will be sent, remove what will not.

Two more ways in:

• Dropping a file on the terminal turns the path the terminal types into an attachment.
• Ctrl+Shift+V (or Alt+Shift+V) pastes an image straight off the system clipboard.

Shuvarie reads each file, sniffs its type and either converts it to markdown (documents) or sends it as multimodal content (images).

Supported formats#

• Documents — pdf, docx, pptx, odt/ods/odp, the xls family, rtf, epub, csv (converted to GitHub-Flavored Markdown), and plain text/code (read line-windowed).
• Images — png, jpeg, gif, webp. Sent as multimodal parts when the model supports them, capped and downscaled to fit.
• Legacy office — .doc/.ppt need an external converter (below).

Attached images render right in the chat pane; the /images command (or a click on the image) opens the fullscreen viewer. The model can also read documents and images through the read_file tool (read_file plan.pdf, read_file shot.png): text files get the line-windowed view, documents return converted markdown, and images return image content to multimodal models.

attachments#

attachments {
    max-images 8        // most images attached to one request (1..=128)
    image-budget 16     // total image budget in MiB (1..=4096)
    image-edge 1568     // long-edge cap in pixels (32..=8000)
    office-converter "soffice"   // external converter for legacy .doc/.ppt
    clipboard-command "wl-paste --type image/png"   // program printing the clipboard image to stdout
}

office-converter names a program (or an absolute path) invoked as --headless --convert-to <docx|pptx> --outdir <dir> <file> — the LibreOffice/soffice CLI. When it is unset, legacy .doc/.ppt files are rejected with a conversion hint.

clipboard-command is the program (with optional arguments) that prints the clipboard image — png, jpeg, gif or webp — to standard output; it is read when you press the paste-image key. When it is unset a platform helper is auto-detected: wl-paste, xclip or pngpaste (and PowerShell on Windows). Pasted images are staged to a temp file under the same size and edge limits as any other image, then enter the ordinary attachment pipeline. Over SSH the clipboard lives on the client, so name a command that ferries it in.

When a drop does nothing#

A dragged file reaches the program only as far as the terminal is willing to send it, and terminals disagree: some paste the path quoted or backslash-escaped, some percent-encode it, some (kitty 0.47+) deliver nothing at all. Set SHUVARIE_TRACE_INPUT=1 (or a file path) to log every raw input event — paste payloads and keys — to $TMPDIR/shuvarie-input-trace.log, which shows what your terminal actually sends. Ctrl+Shift+V is the fallback that does not depend on the terminal at all.

ui.image#

How images are painted (protocol and cell size) is configured under ui.